Scene Recognition Classifiers

COMP3204: Computer Vision

---

Contents

• Introduction
  • Task Description
  • Project Contents
• Running the Application
• Usage
  • Performing Image Classification

---

Introduction

Task Description

• As a group, use the Open-IMAJ library to develop three scene recognition image classifiers.
• The specification of the classifiers is as follows:
  • Classifier 1 : A K-Nearest Neighbour Classifier using the Tiny Image feature as the feature vector for each image.
  • Classifier 2 : A set of 15 one-vs-all Linear Classifiers using a Bag-of-Visual-Words feature based on fixed size densely-sampled pixel patches.
  • Classifier 3 : The best possible Classifier the group can implement (i.e., free to make whatever classifier desired with the goal of maximising performance.)
    • After trialing three different classifiers, the team chose a Naive Bayes Classifier that uses a Pyramid Histogram of Words feature vector based on Dense SIFT features.
• Each of these classifiers must then be trained on provided training data and used to classify provided testing data.
  • The application of each classifier to the testing data is referred to as a Run (i.e., classifier 1 = run 1, classifier 2 = run 2, ...)

Project Contents

Source Code

• Three directories:
  • Run1 : Contains the implementation of the first classifier (K-Nearest Neighbour Classifier).
  • Run2 : Contains the implementation of the second classifier (Linear Classifiers).
  • Run3 : Contains the implementation of the third classifier (best possible classifier).
  • Each directory contains one file for the classifier used in the run, and one file for the feature extractor.
• The following java files:
• App.java : Program used to evaluate the performance of each classifier, and run it on the provided testing data.
• MyClassifier.java and Tuple.java : Helper classes for the classifier implementations.

Documentation

• Documentation.pdf : A description of the group's implementation of all three classifiers.

---

Running the Application

• Only the App.java class is runnable.
• The App.java class defines methods (run1(), run1() and run3() ) that evaluate and run each of the classifiers on the provided training and testing data.
• In the App.java main method, by default, these methods are all called.
• The results of evaluating the classifiers as well as general status messages are outputted to the console.
• The results of classifying the provided testing data with the classifiers are written to text files (run1.txt, run2.txt and run3.txt) as a list of pairs of <image name> <classification>.
• Calls to each of these methods can be removed/commetted out in order to evaluate and run individual classifiers.

---

Usage

Performing Image Classification

• All three of the implemented classifiers extend the MyClassifier.java class, which defines the basic structure and methods of all of the three classifiers.
• One of these methods is makeGuesses():

public ArrayList<Tuple<String, String>> makeGuesses(VFSListDataset<FImage> dataset) { ... }

• After a classifier instance has been instantiated with training data, the makeGuesses() method can be used to annotate a set of unlabelled images.
• The images are passed in as a VFSListDataset, and the method returns an ArrayList of Tuples, where each tuple is a String, String pair, with the first element being the name of the image, and the second being the annotation of this image.

---